From vision to cognition and action
In the last few years computer vision has made tremendous progress in answering questions such as “what is where” in an image. But human abilities to act on and reason about the visual world exceed current systems by far. The first part of the talk will present recent work on the exquisite adaptability of the visual system's active role in acquiring information in one of the most often carried out visual behaviors, i.e. blinking. Subjects were able to learn environmental regularities and adapted their blinking behavior strategically to better detect future events, based on their current beliefs. The observed behavior is in accordance with a computational model based on optimal control trading off intrinsic costs for blink suppression with task-related costs for missing an event under perceptual uncertainty. The second part of the talk will revisit a set of classic cognitive vision tasks, the Bongard problems. Using a formal language representing complex visual concepts and Bayesian inference, complex visual concepts can be induced from the examples that are provided in each Bongard problem, very similar to previous work on the rational analysis of rule-based concept learning. Contrary to other concept learning problems the examples are not random in Bongard problems, instead they are carefully chosen to communicate the concept, hence requiring pragmatic reasoning. Taking pragmatic reasoning into account we find good agreement between the concepts with high posterior probability and the solutions formulated by Bongard himself. While this approach is far from solving all Bongard problems, it solves the biggest fraction yet.
This is joint work with Stefan Depeweg, Frank Jaekel, Stefan Helfmann, and David Hoppe.